onnx.cpp 37.5 KB
Newer Older
Paul's avatar
Paul committed
1
2
3
4
5
6
7
8
#include <google/protobuf/text_format.h>
#include <google/protobuf/io/zero_copy_stream_impl.h>
#include <onnx.pb.h>
#include <iostream>
#include <fstream>
#include <unordered_map>
#include <functional>
#include <array>
Paul's avatar
Paul committed
9
#include <utility>
10
#include <vector>
Paul's avatar
Paul committed
11

Paul's avatar
Paul committed
12
13
14
15
16
17
#include <migraphx/fallthrough.hpp>
#include <migraphx/program.hpp>
#include <migraphx/operators.hpp>
#include <migraphx/ranges.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/config.hpp>
18
#include <migraphx/onnx.hpp>
Paul's avatar
Paul committed
19
20

namespace migraphx {
Paul's avatar
Paul committed
21
inline namespace MIGRAPHX_INLINE_NS {
Paul's avatar
Paul committed
22
23
24
25
26

struct onnx_parser
{
    using attribute_map = std::unordered_map<std::string, onnx::AttributeProto>;
    using node_map      = std::unordered_map<std::string, onnx::NodeProto>;
Paul's avatar
Paul committed
27
    using op_func = std::function<instruction_ref(attribute_map, std::vector<instruction_ref>)>;
Paul's avatar
Paul committed
28
29
    node_map nodes;
    std::unordered_map<std::string, instruction_ref> instructions;
Scott Thornton's avatar
Scott Thornton committed
30
    program prog    = program();
31
    bool is_pytorch = false;
Paul's avatar
Paul committed
32
33

    std::unordered_map<std::string, op_func> ops;
34
    std::unordered_map<std::string, operation> actv_funcs;
Paul's avatar
Paul committed
35
36
37

    onnx_parser()
    {
Shucai Xiao's avatar
Shucai Xiao committed
38
        add_generic_op("MatMul", op::dot{});
Khalique's avatar
Khalique committed
39
        add_generic_op("Relu", op::relu{});
Khalique's avatar
Khalique committed
40
41
        add_generic_op("Sigmoid", op::sigmoid{});
        add_generic_op("Abs", op::abs{});
Shucai Xiao's avatar
Shucai Xiao committed
42
43
        add_generic_op("Exp", op::exp{});
        add_generic_op("Log", op::log{});
Khalique's avatar
Khalique committed
44
45
        // disable dropout for inference
        add_generic_op("Dropout", op::identity{});
Khalique's avatar
Khalique committed
46
        add_generic_op("Identity", op::identity{});
Shucai Xiao's avatar
Shucai Xiao committed
47
48
49
        add_generic_op("Sin", op::sin{});
        add_generic_op("Cos", op::cos{});
        add_generic_op("Tan", op::tan{});
50
51
        add_generic_op("Sinh", op::sinh{});
        add_generic_op("Cosh", op::cosh{});
52
        add_generic_op("Tanh", op::tanh{});
53
54
55
        add_generic_op("Asin", op::asin{});
        add_generic_op("Acos", op::acos{});
        add_generic_op("Atan", op::atan{});
Paul's avatar
Paul committed
56

Khalique's avatar
Khalique committed
57
58
59
60
61
        add_binary_op("Add", op::add{});
        add_binary_op("Div", op::div{});
        add_binary_op("Mul", op::mul{});
        add_binary_op("Sub", op::sub{});

Khalique's avatar
Khalique committed
62
63
64
        add_variadic_op("Sum", op::add{});
        add_variadic_op("Max", op::max{});
        add_variadic_op("Min", op::min{});
Paul's avatar
Paul committed
65

Khalique's avatar
Khalique committed
66
        add_mem_op("ImageScaler", &onnx_parser::parse_imagescaler);
67
        add_mem_op("LeakyRelu", &onnx_parser::parse_leaky_relu);
Khalique's avatar
Khalique committed
68
        add_mem_op("Elu", &onnx_parser::parse_elu);
Paul's avatar
Paul committed
69
70
        add_mem_op("Constant", &onnx_parser::parse_constant);
        add_mem_op("Conv", &onnx_parser::parse_conv);
Paul's avatar
Paul committed
71
72
        add_mem_op("MaxPool", &onnx_parser::parse_pooling);
        add_mem_op("AveragePool", &onnx_parser::parse_pooling);
73
74
        add_mem_op("GlobalMaxPool", &onnx_parser::parse_pooling);
        add_mem_op("GlobalAveragePool", &onnx_parser::parse_pooling);
Paul's avatar
Paul committed
75
        add_mem_op("Reshape", &onnx_parser::parse_reshape);
Paul's avatar
Paul committed
76
77
        add_mem_op("Flatten", &onnx_parser::parse_flatten);
        add_mem_op("Gemm", &onnx_parser::parse_gemm);
78
        add_mem_op("BatchNormalization", &onnx_parser::parse_batchnorm);
Paul's avatar
Paul committed
79
        add_mem_op("Softmax", &onnx_parser::parse_softmax);
80
81
82
        add_mem_op("Squeeze", &onnx_parser::parse_squeeze);
        add_mem_op("Unsqueeze", &onnx_parser::parse_unsqueeze);
        add_mem_op("Slice", &onnx_parser::parse_slice);
Scott Thornton's avatar
Scott Thornton committed
83
        add_mem_op("Concat", &onnx_parser::parse_concat);
84
85
86
        add_mem_op("Gather", &onnx_parser::parse_gather);
        add_mem_op("Shape", &onnx_parser::parse_shape);
        add_mem_op("ConstantFill", &onnx_parser::parse_constant_fill);
Khalique's avatar
Khalique committed
87
        add_mem_op("Transpose", &onnx_parser::parse_transpose);
Shucai Xiao's avatar
Shucai Xiao committed
88
        add_mem_op("RNN", &onnx_parser::parse_rnn);
89
90
91
92
93
94
95
96
97
98

        // init the activation function map
        init_actv_func();
    }

    void init_actv_func()
    {
        actv_funcs.insert(std::make_pair("tanh", op::tanh{}));
        actv_funcs.insert(std::make_pair("relu", op::relu{}));
        actv_funcs.insert(std::make_pair("sigmoid", op::sigmoid{}));
Paul's avatar
Paul committed
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
    }

    template <class F>
    void add_op(std::string name, F f)
    {
        ops.emplace(name, f);
    }

    template <class F>
    void add_mem_op(std::string name, F f)
    {
        ops.emplace(name, [=](auto&&... xs) {
            return std::mem_fn(f)(*this, name, std::forward<decltype(xs)>(xs)...);
        });
    }
Khalique's avatar
Khalique committed
114

115
    template <class T>
Khalique's avatar
Khalique committed
116
    void add_binary_op(std::string name, T x)
117
118
    {
        ops.emplace(name, [this, x](attribute_map attributes, std::vector<instruction_ref> args) {
Scott Thornton's avatar
Scott Thornton committed
119
            if(args.size() != 2)
Paul's avatar
Paul committed
120
                MIGRAPHX_THROW("binary operators should have 2 operands");
121
122
123
124
125
126
127
128
129
130
131
132
            if(contains(attributes, "broadcast"))
            {
                uint64_t broadcasted = parse_value(attributes.at("broadcast")).at<uint64_t>();
                if(broadcasted != 0)
                {
                    uint64_t axis = (contains(attributes, "axis"))
                                        ? parse_value(attributes.at("axis")).at<uint64_t>()
                                        : 0;
                    auto l =
                        prog.add_instruction(op::broadcast{axis, args[0]->get_shape()}, args[1]);
                    return prog.add_instruction(x, args[0], l);
                }
133
                return prog.add_instruction(x, args);
134
            }
Paul's avatar
Paul committed
135
            else
136
            {
Khalique's avatar
Khalique committed
137
                return add_broadcastable_binary_op(args[0], args[1], x);
138
139
140
141
            }
        });
    }

Khalique's avatar
Khalique committed
142
143
144
145
146
    template <class T>
    instruction_ref add_broadcastable_binary_op(instruction_ref arg0, instruction_ref arg1, T x)
    {
        if(arg0->get_shape() != arg1->get_shape())
        {
Khalique's avatar
Khalique committed
147
148
149
150
151
152
153
154
155
156
157
158
159
            // Example:
            // s0 = (3,2,4,5) and s1 = (2,1,1)
            //
            // In this case we need to broadcast (:,1,1) portion of
            // s1 plus broadcast the 1st dimension of s1
            // giving output_lens = (3,2,4,5)
            //
            // Another example:
            // s0 = (3,2,1,5) and s1 = (2,7,5)
            // In this case we need to broadcast the (:,:,1:,:) axis
            // of s0 plus the 1st dimension of s1 giving
            // output_lens = (3,2,7,5)
            //
Khalique's avatar
Khalique committed
160
161
162
163
164
165
166
167
            // Get lengths for both arguments
            const std::vector<std::size_t>* s0 = &arg0->get_shape().lens();
            const std::vector<std::size_t>* s1 = &arg1->get_shape().lens();

            // Make sure s0 is the smaller size
            if(s0->size() > s1->size())
                std::swap(s0, s1);

Khalique's avatar
Khalique committed
168
            std::vector<std::size_t> output_lens(*s1);
Khalique's avatar
Khalique committed
169
170
            auto offset = s1->size() - s0->size();
            std::transform(s0->begin(),
Khalique's avatar
Khalique committed
171
172
173
174
                           s0->end(),
                           s1->begin() + offset,
                           output_lens.begin() + offset,
                           [](auto a, auto b) { return std::max(a, b); });
Khalique's avatar
Khalique committed
175
176
177
178
179
180
181
182
183

            auto l0 = prog.add_instruction(op::multibroadcast{output_lens}, arg0);
            auto l1 = prog.add_instruction(op::multibroadcast{output_lens}, arg1);
            return prog.add_instruction(x, l0, l1);
        }
        else
        {
            return prog.add_instruction(x, {arg0, arg1});
        }
184
185
    }

Paul's avatar
Paul committed
186
    template <class T>
Paul's avatar
Paul committed
187
188
    void add_generic_op(std::string name, T x)
    {
Paul's avatar
Paul committed
189
        ops.emplace(name, [this, x](attribute_map, std::vector<instruction_ref> args) {
Paul's avatar
Paul committed
190
191
192
193
            return prog.add_instruction(x, args);
        });
    }

Khalique's avatar
Khalique committed
194
    template <class T>
Khalique's avatar
Khalique committed
195
    void add_variadic_op(std::string name, T x)
Khalique's avatar
Khalique committed
196
    {
Khalique's avatar
Khalique committed
197
198
        ops.emplace(name, [this, x](attribute_map, std::vector<instruction_ref> args) {
            return std::accumulate(std::next(args.begin()),
Khalique's avatar
Khalique committed
199
200
201
202
203
                                   args.end(),
                                   args.front(),
                                   [this, x](instruction_ref a, instruction_ref b) {
                                       return add_broadcastable_binary_op(a, b, x);
                                   });
Khalique's avatar
Khalique committed
204
        });
Khalique's avatar
Khalique committed
205
206
    }

Paul's avatar
Paul committed
207
    instruction_ref
Paul's avatar
Paul committed
208
    parse_softmax(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
209
210
    {
        auto dims = args.front()->get_shape().lens();
Scott Thornton's avatar
Scott Thornton committed
211
212
        auto r =
            prog.add_instruction(op::reshape{{long(dims[0]), long(dims[1]), 1, 1}}, args.front());
213
214
        auto s = prog.add_instruction(op::softmax{}, r);
        return prog.add_instruction(op::reshape{{long(dims[0]), long(dims[1])}}, s);
Paul's avatar
Paul committed
215
216
    }

Paul's avatar
Paul committed
217
    instruction_ref
Paul's avatar
Paul committed
218
    parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
219
    {
220
        op::convolution op;
Paul's avatar
Paul committed
221
222
        if(contains(attributes, "pads"))
        {
Scott Thornton's avatar
Scott Thornton committed
223
            if(contains(attributes, "auto_pad"))
224
            {
Paul's avatar
Paul committed
225
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
226
227
228
            }
            std::vector<std::size_t> padding(4);
            copy(attributes["pads"].ints(), padding.begin());
Scott Thornton's avatar
Scott Thornton committed
229
            if(padding.size() != 4)
230
            {
Paul's avatar
Paul committed
231
                MIGRAPHX_THROW("padding should have 4 values");
232
            }
Scott Thornton's avatar
Scott Thornton committed
233
            if(padding[0] != padding[2] || padding[1] != padding[3])
234
            {
Paul's avatar
Paul committed
235
                MIGRAPHX_THROW("migraphx does not support asymetric padding");
236
237
238
            }
            op.padding[0] = padding[0];
            op.padding[1] = padding[1];
Paul's avatar
Paul committed
239
        }
Paul's avatar
Paul committed
240
241
242
243
244
245
246
247
        if(contains(attributes, "strides"))
        {
            copy(attributes["strides"].ints(), op.stride.begin());
        }
        if(contains(attributes, "dilations"))
        {
            copy(attributes["dilations"].ints(), op.dilation.begin());
        }
Scott Thornton's avatar
Scott Thornton committed
248
        if(contains(attributes, "auto_pad"))
249
250
        {
            auto s = attributes["auto_pad"].s();
Scott Thornton's avatar
Scott Thornton committed
251
            if(contains(attributes, "pads") and to_upper(s) != "NOTSET")
252
            {
Paul's avatar
Paul committed
253
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
254
255
            }

wsttiger's avatar
fixes  
wsttiger committed
256
            if(s.find("SAME") != std::string::npos)
257
258
259
260
            {
                op.padding_mode = op::convolution::same;
            }
        }
Khalique's avatar
Khalique committed
261
262
263
264
        if(contains(attributes, "group"))
        {
            op.group = parse_value(attributes.at("group")).at<int>();
        }
Paul's avatar
Paul committed
265
266
267
268
        if(args.size() == 3)
        {
            uint64_t axis = 1;
            auto l1       = prog.add_instruction(op, args[0], args[1]);
Scott Thornton's avatar
Scott Thornton committed
269
            auto l2       = prog.add_instruction(op::broadcast{axis, l1->get_shape()}, args[2]);
270
            return prog.add_instruction(op::add{}, l1, l2);
Paul's avatar
Paul committed
271
        }
Paul's avatar
Paul committed
272
273
        return prog.add_instruction(op, args);
    }
Paul's avatar
Paul committed
274

Paul's avatar
Paul committed
275
276
277
    instruction_ref parse_pooling(const std::string& name,
                                  attribute_map attributes,
                                  std::vector<instruction_ref> args)
Paul's avatar
Paul committed
278
    {
Khalique's avatar
Khalique committed
279
280
        op::pooling op{ends_with(name, "MaxPool") ? "max" : "average"};
        if(starts_with(name, "Global"))
281
        {
Khalique's avatar
Khalique committed
282
283
            auto lens  = args.front()->get_shape().lens();
            op.lengths = {lens[2], lens[3]};
284
        }
Paul's avatar
Paul committed
285
286
        if(contains(attributes, "pads"))
        {
287
288
            std::vector<std::size_t> padding(4);
            copy(attributes["pads"].ints(), padding.begin());
Scott Thornton's avatar
Scott Thornton committed
289
            if(padding.size() != 4)
290
            {
Paul's avatar
Paul committed
291
                MIGRAPHX_THROW("padding should have 4 values");
292
            }
Scott Thornton's avatar
Scott Thornton committed
293
            if(padding[0] != padding[2] || padding[1] != padding[3])
294
            {
Paul's avatar
Paul committed
295
                MIGRAPHX_THROW("migraphx does not support asymetric padding");
296
297
298
            }
            op.padding[0] = padding[0];
            op.padding[1] = padding[1];
Paul's avatar
Paul committed
299
300
301
302
303
304
305
306
307
        }
        if(contains(attributes, "strides"))
        {
            copy(attributes["strides"].ints(), op.stride.begin());
        }
        if(contains(attributes, "kernel_shape"))
        {
            copy(attributes["kernel_shape"].ints(), op.lengths.begin());
        }
Scott Thornton's avatar
Scott Thornton committed
308
        if(contains(attributes, "auto_pad"))
309
310
        {
            auto s = attributes["auto_pad"].s();
Scott Thornton's avatar
Scott Thornton committed
311
            if(to_upper(s) != "NOTSET")
312
            {
Paul's avatar
Paul committed
313
                MIGRAPHX_THROW("auto_pad is not supported for pooling");
314
315
316
            }
        }

Paul's avatar
Paul committed
317
        return prog.add_instruction(op, std::move(args));
Paul's avatar
Paul committed
318
319
    }

Paul's avatar
Paul committed
320
    instruction_ref
Paul's avatar
Paul committed
321
    parse_reshape(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
322
    {
323
        op::reshape op;
Paul's avatar
Paul committed
324
325
326
327
328
329
330
        if(args.size() == 1)
        {
            literal s = parse_value(attributes.at("shape"));
            s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
        }
        if(args.size() == 2)
        {
Paul's avatar
Paul committed
331
            literal s = args[1]->get_literal();
Paul's avatar
Paul committed
332
            s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
Paul's avatar
Paul committed
333
        }
Paul's avatar
Paul committed
334
335
336
        return prog.add_instruction(op, args[0]);
    }

Paul's avatar
Paul committed
337
    instruction_ref
Paul's avatar
Paul committed
338
    parse_flatten(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
339
    {
340
        uint64_t axis = 1;
Paul's avatar
Paul committed
341
342
343
344
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
345
        return prog.add_instruction(op::flatten{axis}, args[0]);
Paul's avatar
Paul committed
346
347
    }

348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
    instruction_ref
    parse_squeeze(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::squeeze op;
        literal s = parse_value(attributes.at("axes"));
        s.visit([&](auto v) { copy(v, std::back_inserter(op.axes)); });
        return prog.add_instruction(op, args[0]);
    }

    instruction_ref
    parse_unsqueeze(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::unsqueeze op;
        literal s = parse_value(attributes.at("axes"));
        s.visit([&](auto v) { copy(v, std::back_inserter(op.axes)); });
        return prog.add_instruction(op, args[0]);
    }

Scott Thornton's avatar
Scott Thornton committed
366
367
368
369
370
371
372
    instruction_ref
    parse_concat(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        std::size_t axis = parse_value(attributes.at("axis")).at<int>();
        op::concat op{axis};
        return prog.add_instruction(op, std::move(args));
    }
373

374
375
376
377
378
379
380
381
    instruction_ref
    parse_gather(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        std::size_t axis = 0;
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
382
        op::gather op{axis};
383
384
385
        return prog.add_instruction(op, std::move(args));
    }

386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
    instruction_ref
    parse_slice(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::slice op;
        if(contains(attributes, "axes"))
        {
            literal s = parse_value(attributes.at("axes"));
            s.visit([&](auto v) { copy(v, std::back_inserter(op.axes)); });
        }
        {
            literal s = parse_value(attributes.at("ends"));
            s.visit([&](auto v) { copy(v, std::back_inserter(op.ends)); });
        }
        {
            literal s = parse_value(attributes.at("starts"));
            s.visit([&](auto v) { copy(v, std::back_inserter(op.starts)); });
        }
        return prog.add_instruction(op, args[0]);
    }

Paul's avatar
Paul committed
406
407
408
    instruction_ref parse_constant(const std::string&,
                                   attribute_map attributes,
                                   const std::vector<instruction_ref>&)
Paul's avatar
Paul committed
409
410
411
412
    {
        literal v = parse_value(attributes.at("value"));
        return prog.add_literal(v);
    }
Paul's avatar
Paul committed
413

Paul's avatar
Paul committed
414
    instruction_ref
Paul's avatar
Paul committed
415
    parse_gemm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
416
417
    {
        float alpha = 1.0f;
Khalique's avatar
Khalique committed
418
        float beta  = 1.0f;
Paul's avatar
Paul committed
419
420
421
422
423
424
425
426
        bool transa = false;
        bool transb = false;
        if(contains(attributes, "alpha"))
        {
            alpha = parse_value(attributes.at("alpha")).at<float>();
        }
        if(contains(attributes, "beta"))
        {
427
            beta = parse_value(attributes.at("beta")).at<float>();
Paul's avatar
Paul committed
428
429
430
431
432
433
434
435
436
437
        }
        if(contains(attributes, "transA"))
        {
            transa = parse_value(attributes.at("transA")).at<bool>();
        }
        if(contains(attributes, "transB"))
        {
            transb = parse_value(attributes.at("transB")).at<bool>();
        }
        std::vector<int64_t> perm = {1, 0};
438
439
        auto l1 = (transa) ? prog.add_instruction(op::transpose{perm}, args[0]) : args[0];
        auto l2 = (transb) ? prog.add_instruction(op::transpose{perm}, args[1]) : args[1];
Paul's avatar
Paul committed
440
441
        if(args.size() == 3)
        {
Khalique's avatar
Khalique committed
442
            if(beta != 0.f)
443
            {
Khalique's avatar
Khalique committed
444
                auto l3 = prog.add_instruction(op::dot{alpha}, l1, l2);
Khalique's avatar
Khalique committed
445
                auto l4 = args[2];
Khalique's avatar
Khalique committed
446
                if(l4->get_shape().scalar()) // ignore args[2] (no C value added to alpha*A*B)
Khalique's avatar
Khalique committed
447
                    return l3;
Khalique's avatar
Khalique committed
448
                if(beta != 1.f)
Khalique's avatar
Khalique committed
449
450
                {
                    auto beta_val = prog.add_literal(beta);
Khalique's avatar
Khalique committed
451
452
                    auto l5 = prog.add_instruction(op::scalar{args[2]->get_shape()}, beta_val);
                    l4      = prog.add_instruction(op::mul{}, args[2], l5);
Khalique's avatar
Khalique committed
453
454
                }
                return add_broadcastable_binary_op(l3, l4, op::add{});
455
            }
Paul's avatar
Paul committed
456
        }
Shucai Xiao's avatar
Shucai Xiao committed
457
        return prog.add_instruction(op::dot{alpha, beta}, l1, l2);
Paul's avatar
Paul committed
458
459
    }

460
    instruction_ref
Paul's avatar
Paul committed
461
    parse_batchnorm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
462
    {
Scott Thornton's avatar
Scott Thornton committed
463
464
        float epsilon                                     = 1e-5f;
        float momentum                                    = 0.9f;
465
        op::batch_norm_inference::bn_infer_mode_t bn_mode = op::batch_norm_inference::spatial;
Scott Thornton's avatar
Scott Thornton committed
466
        bool is_test                                      = false;
467
468
469
470
471
472
        if(contains(attributes, "epsilon"))
        {
            epsilon = parse_value(attributes.at("epsilon")).at<float>();
        }
        if(contains(attributes, "momentum"))
        {
473
            momentum = parse_value(attributes.at("momentum")).at<float>();
474
475
476
        }
        if(contains(attributes, "is_test"))
        {
wsttiger's avatar
wsttiger committed
477
            is_test = parse_value(attributes.at("is_test")).at<uint64_t>() > 0;
478
479
480
        }
        if(contains(attributes, "spatial"))
        {
481
            bn_mode = (parse_value(attributes.at("spatial")).at<uint64_t>() > 0)
482
483
                          ? op::batch_norm_inference::spatial
                          : op::batch_norm_inference::per_activation;
484
        }
Paul's avatar
Paul committed
485
        (void)is_test;
Paul's avatar
Paul committed
486
        op::batch_norm_inference op{epsilon, momentum, bn_mode};
Paul's avatar
Paul committed
487
        return prog.add_instruction(op, std::move(args));
488
489
    }

490
491
492
493
    instruction_ref parse_leaky_relu(const std::string&,
                                     attribute_map attributes,
                                     std::vector<instruction_ref> args)
    {
Khalique's avatar
Khalique committed
494
        float alpha = 0.01; // default alpha val for leaky relu
495
496
497
498
499
500
501
502
        if(contains(attributes, "alpha"))
        {
            alpha = parse_value(attributes.at("alpha")).at<float>();
        }
        op::leaky_relu op{alpha};
        return prog.add_instruction(op, args.front());
    }

Khalique's avatar
Khalique committed
503
504
    instruction_ref
    parse_elu(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
505
506
507
508
509
510
511
512
513
514
    {
        float alpha = 1.0; // default alpha val for elu
        if(contains(attributes, "alpha"))
        {
            alpha = parse_value(attributes.at("alpha")).at<float>();
        }
        op::elu op{alpha};
        return prog.add_instruction(op, args.front());
    }

Khalique's avatar
Khalique committed
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
    instruction_ref parse_imagescaler(const std::string&,
                                      attribute_map attributes,
                                      std::vector<instruction_ref> args)
    {
        float scale = 1.0;
        std::vector<float> bias{};
        if(contains(attributes, "scale"))
        {
            scale = parse_value(attributes.at("scale")).at<float>();
        }

        if(contains(attributes, "bias"))
        {
            auto&& bias_floats = attributes["bias"].floats();
            bias               = std::vector<float>(bias_floats.begin(), bias_floats.end());
        }
        auto input_shape = args.front()->get_shape();
Khalique's avatar
Khalique committed
532

Khalique's avatar
Khalique committed
533
534
        auto scale_val = prog.add_literal(scale);
        auto bias_vals = prog.add_literal(
Paul's avatar
Paul committed
535
            migraphx::literal{migraphx::shape{migraphx::shape::float_type, {bias.size()}}, bias});
Khalique's avatar
Khalique committed
536

Paul's avatar
Paul committed
537
538
        auto scale_tensor = prog.add_instruction(migraphx::op::scalar{input_shape}, scale_val);
        auto img_scaled   = prog.add_instruction(migraphx::op::mul{}, args.front(), scale_tensor);
Paul's avatar
Paul committed
539
        auto bias_bcast = prog.add_instruction(migraphx::op::broadcast{1, input_shape}, bias_vals);
Paul's avatar
Paul committed
540
        return prog.add_instruction(migraphx::op::add{}, img_scaled, bias_bcast);
Khalique's avatar
Khalique committed
541
    }
Khalique's avatar
Khalique committed
542

Khalique's avatar
Khalique committed
543
544
    instruction_ref
    parse_transpose(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
545
546
547
548
549
550
551
    {
        std::vector<int64_t> perm{};
        if(contains(attributes, "perm"))
        {
            auto&& perm_vals = attributes["perm"].ints();
            perm             = std::vector<int64_t>(perm_vals.begin(), perm_vals.end());
        }
Paul's avatar
Paul committed
552
        return prog.add_instruction(migraphx::op::transpose{perm}, args.front());
Khalique's avatar
Khalique committed
553
554
    }

555
556
557
    // Use a literal instruction to replace the shape since, output of
    // shape operator are literals in migraphx
    instruction_ref
Shucai Xiao's avatar
Shucai Xiao committed
558
    parse_shape(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
559
560
    {
        if(args.size() != 1)
561
            MIGRAPHX_THROW("Shape: operator should have 1 operand");
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
        std::vector<std::size_t> arg_shape = args[0]->get_shape().lens();
        std::vector<int64_t> vec_shape(arg_shape.size());
        migraphx::shape s(migraphx::shape::int64_type, {arg_shape.size()});
        std::transform(arg_shape.begin(), arg_shape.end(), vec_shape.begin(), [](auto i) {
            return int64_t(i);
        });
        return prog.add_literal(migraphx::literal{s, vec_shape});
    }

    // Use a literal instruction to replace the constantFill operator. In RNN, input shape
    // and value are fixed, so no need to do the actual computation for the constantFill
    // operator
    instruction_ref parse_constant_fill(const std::string&,
                                        attribute_map attributes,
                                        std::vector<instruction_ref> args)
    {
        int input_as_shape = 0;
        int dtype          = 1;
        float value        = 0.0f;

        if(contains(attributes, "dtype"))
        {
            dtype = parse_value(attributes.at("dtype")).at<int>();
        }
        migraphx::shape::type_t type = get_type(dtype);

        if(contains(attributes, "input_as_shape"))
        {
            input_as_shape = parse_value(attributes.at("input_as_shape")).at<int>();
        }

        if(contains(attributes, "value"))
        {
            value = parse_value(attributes.at("value")).at<float>();
        }

Shucai Xiao's avatar
Shucai Xiao committed
598
599
        if(contains(attributes, "extra_shape"))
        {
600
            MIGRAPHX_THROW("ConstantFill: cannot handle extra shape attribute");
601
602
        }

603
604
        if(input_as_shape == 1)
        {
Shucai Xiao's avatar
Shucai Xiao committed
605
            if(args.size() != 1)
606
            {
607
                MIGRAPHX_THROW("ConstantFill: need an input argument as output shape");
608
609
            }

Shucai Xiao's avatar
Shucai Xiao committed
610
611
            if(contains(attributes, "shape"))
            {
612
                MIGRAPHX_THROW("ConstantFill: cannot set the shape argument and pass in an input "
Shucai Xiao's avatar
Shucai Xiao committed
613
                               "at the same time");
614
615
            }

616
617
618
            migraphx::argument in = args[0]->eval();
            if(in.empty())
            {
619
                MIGRAPHX_THROW("ConstantFill: cannot handle dynamic shape as input");
620
            }
621

622
623
624
            std::vector<std::size_t> dims;
            in.visit([&](auto input) { dims.assign(input.begin(), input.end()); });
            migraphx::shape s(type, dims);
625
626
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
627
628
629
        }
        else if(input_as_shape == 0)
        {
Shucai Xiao's avatar
Shucai Xiao committed
630
631
            if(!contains(attributes, "shape"))
            {
632
                MIGRAPHX_THROW("ConstantFill: attribute output shape is needed");
633
634
635
            }

            literal ls = parse_value(attributes.at("shape"));
636
            std::vector<std::size_t> dims;
Shucai Xiao's avatar
Shucai Xiao committed
637
            ls.visit([&](auto s) { dims.assign(s.begin(), s.end()); });
638
            migraphx::shape s{type, dims};
639
640
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
641
642
643
        }
        else
        {
644
            MIGRAPHX_THROW("ConstantFill: wrong value of attribute input_as_shape");
645
646
647
        }
    }

Shucai Xiao's avatar
Shucai Xiao committed
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
    instruction_ref
    parse_rnn(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        migraphx::shape input_shape = args[0]->get_shape();
        migraphx::shape w_shape     = args[1]->get_shape();
        std::size_t hidden_size     = w_shape.lens()[1];

        if(contains(attributes, "hidden_size"))
        {
            hidden_size = parse_value(attributes.at("hidden_size")).at<int>();
        }
        else
        {
            MIGRAPHX_THROW("RNN: hidden size attribute missing");
        }

        std::string activation_func = {"tanh"};
        if(contains(attributes, "activations"))
        {
            activation_func = attributes.at("activations").strings(0);
        }

670
        if(actv_funcs.count(activation_func) == 0)
Shucai Xiao's avatar
Shucai Xiao committed
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
        {
            MIGRAPHX_THROW("RNN: activation function " + activation_func + " not supported");
        }

        // Handling of direction to be added later
        std::string direction{"forward"};
        if(contains(attributes, "direction"))
        {
            direction = attributes.at("direction").s();
        }

        op::rnn::rnn_direction_t dirct = op::rnn::forward;
        if(direction == "bidirectional")
        {
            dirct = op::rnn::bidirectional;
        }
        else if(direction == "reverse")
        {
            dirct = op::rnn::reverse;
        }

        // To be added later
        float clip = 0.0;
        if(contains(attributes, "clip"))
        {
            clip = parse_value(attributes.at("clip")).at<float>();
        }

699
700
        return prog.add_instruction(op::rnn{hidden_size, actv_funcs[activation_func], dirct, clip},
                                    std::move(args));
Shucai Xiao's avatar
Shucai Xiao committed
701
702
    }

Paul's avatar
Paul committed
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
    void parse_from(std::istream& is)
    {
        onnx::ModelProto model;
        if(model.ParseFromIstream(&is))
        {
            if(model.has_graph())
            {
                this->parse_graph(model.graph());
            }
        }
        else
        {
            throw std::runtime_error("Failed reading");
        }
    }

    void parse_graph(const onnx::GraphProto& graph)
    {
        nodes = get_nodes(graph);
722
723
724
725
726
        std::unordered_map<std::string, onnx::TensorProto> initializer_data;
        for(auto&& f : graph.initializer())
        {
            initializer_data[f.name()] = f;
        }
Paul's avatar
Paul committed
727
728
729
        for(auto&& input : graph.input())
        {
            const std::string& name = input.name();
730
731
732
733
734
735
736
737
738
739
740
741
            // Does the input have an initializer?
            if(contains(initializer_data, name))
            {
                auto t             = initializer_data[name];
                instructions[name] = prog.add_literal(parse_tensor(t));
            }
            else
            {
                // TODO: Get shape of input parameter
                shape s            = parse_type(input.type());
                instructions[name] = prog.add_parameter(name, s);
            }
Paul's avatar
Paul committed
742
743
744
        }
        for(auto&& p : nodes)
        {
745
            this->parse_node(get_name(p.second));
Paul's avatar
Paul committed
746
747
748
        }
    }

Paul's avatar
Paul committed
749
    void parse_node(const std::string& name)
Paul's avatar
Paul committed
750
    {
Paul's avatar
Paul committed
751
        if(name.empty())
Paul's avatar
Paul committed
752
            MIGRAPHX_THROW("Onnx node must have a name");
Paul's avatar
Paul committed
753
754
755
756
757
758
        if(instructions.count(name) == 0)
        {
            auto&& node = nodes.at(name);
            std::vector<instruction_ref> args;
            for(auto&& input : node.input())
            {
759
760
761
762
763
764
765
766
767
768
                // For RNN, LSTM, and GRU operators, one of the input arguments
                // is prim::Undefined, and it is ignored by protobuf. We use a
                // hack to ignore this argument for these three operators
                std::string op_type = node.op_type();
                if((op_type == "RNN" || op_type == "LSTM" || op_type == "GRU") &&
                   input.empty() == true)
                {
                    continue;
                }

Paul's avatar
Paul committed
769
770
                if(nodes.count(input) > 0)
                {
771
                    auto&& iname = get_name(nodes.at(input));
Paul's avatar
Paul committed
772
                    assert(name != iname);
Paul's avatar
Paul committed
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
                    this->parse_node(iname);
                    args.push_back(instructions.at(iname));
                }
                else
                {
                    args.push_back(instructions.at(input));
                }
            }
            if(ops.count(node.op_type()) == 0)
            {
                instructions[name] = prog.add_instruction(unknown{node.op_type()}, args);
            }
            else
            {
                instructions[name] = ops[node.op_type()](get_attributes(node), args);
            }
        }
    }

    static attribute_map get_attributes(const onnx::NodeProto& node)
    {
        std::unordered_map<std::string, onnx::AttributeProto> result;
        for(auto&& attr : node.attribute())
        {
            result[attr.name()] = attr;
        }
        return result;
    }

802
803
804
805
    static std::string get_name(const onnx::NodeProto& node)
    {
        if(node.name().empty())
        {
Paul's avatar
Paul committed
806
            std::string generated = "migraphx_unnamed_node";
Paul's avatar
Paul committed
807
808
809
810
            return std::accumulate(node.output().begin(),
                                   node.output().end(),
                                   generated,
                                   [](auto x, auto y) { return x + "_" + y; });
811
812
813
814
        }
        return node.name();
    }

Paul's avatar
Paul committed
815
816
817
818
819
    static node_map get_nodes(const onnx::GraphProto& graph)
    {
        std::unordered_map<std::string, onnx::NodeProto> result;
        for(auto&& node : graph.node())
        {
820
            result[get_name(node)] = node;
Paul's avatar
Paul committed
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
            for(auto&& output : node.output())
            {
                result[output] = node;
            }
        }
        return result;
    }

    template <class T>
    static literal from_repeated(shape::type_t t, const T& r)
    {
        std::size_t size = r.size();
        return literal{{t, {size}}, r.begin(), r.end()};
    }

    static literal parse_value(const onnx::AttributeProto& attr)
    {
        switch(attr.type())
        {
        case onnx::AttributeProto::UNDEFINED: return {};
        case onnx::AttributeProto::FLOAT: return literal{attr.f()};
        case onnx::AttributeProto::INT: return literal{attr.i()};
        case onnx::AttributeProto::STRING: return {};
        case onnx::AttributeProto::TENSOR: return parse_tensor(attr.t());
        case onnx::AttributeProto::GRAPH: return {};
Paul's avatar
Paul committed
846
        case onnx::AttributeProto::FLOATS: return from_repeated(shape::float_type, attr.floats());
Paul's avatar
Paul committed
847
848
849
850
851
        case onnx::AttributeProto::INTS: return from_repeated(shape::int64_type, attr.ints());
        case onnx::AttributeProto::STRINGS: return {};
        case onnx::AttributeProto::TENSORS: return {};
        case onnx::AttributeProto::GRAPHS: return {};
        }
Paul's avatar
Paul committed
852
        MIGRAPHX_THROW("Invalid attribute type");
Paul's avatar
Paul committed
853
854
855
856
857
    }

    static literal parse_tensor(const onnx::TensorProto& t)
    {
        std::vector<std::size_t> dims(t.dims().begin(), t.dims().end());
Khalique's avatar
Khalique committed
858
        // in case of scalar constants in onnx file, use dims=1 to fill initializer data
859
        if(dims.empty())
Khalique's avatar
Khalique committed
860
861
862
        {
            dims = {1};
        }
863
864
        if(t.has_raw_data())
        {
wsttiger's avatar
wsttiger committed
865
            const std::string& s = t.raw_data();
Scott Thornton's avatar
Scott Thornton committed
866
867
868
869
870
871
872
873
874
875
876
877
            switch(t.data_type())
            {
            case onnx::TensorProto::UNDEFINED: throw std::runtime_error("");
            case onnx::TensorProto::FLOAT: return literal{{shape::float_type, dims}, s.data()};
            case onnx::TensorProto::UINT8: throw std::runtime_error("");
            case onnx::TensorProto::INT8: return literal{{shape::int32_type, dims}, s.data()};
            case onnx::TensorProto::UINT16: return literal{{shape::int32_type, dims}, s.data()};
            case onnx::TensorProto::INT16: return literal{{shape::int32_type, dims}, s.data()};
            case onnx::TensorProto::INT32: return literal{{shape::int32_type, dims}, s.data()};
            case onnx::TensorProto::INT64: return literal{{shape::int64_type, dims}, s.data()};
            case onnx::TensorProto::STRING: throw std::runtime_error("");
            case onnx::TensorProto::BOOL: return literal{{shape::int32_type, dims}, s.data()};
Paul's avatar
Paul committed
878
            case onnx::TensorProto::FLOAT16: return literal{{shape::half_type, dims}, s.data()};
Scott Thornton's avatar
Scott Thornton committed
879
880
881
882
883
884
            case onnx::TensorProto::DOUBLE: return literal{{shape::double_type, dims}, s.data()};
            case onnx::TensorProto::UINT32: throw std::runtime_error("");
            case onnx::TensorProto::UINT64: throw std::runtime_error("");
            case onnx::TensorProto::COMPLEX64: throw std::runtime_error("");
            case onnx::TensorProto::COMPLEX128: throw std::runtime_error("");
            }
Paul's avatar
Paul committed
885
            MIGRAPHX_THROW("Invalid tensor type");
886
        }
Paul's avatar
Paul committed
887
888
889
890
        switch(t.data_type())
        {
        case onnx::TensorProto::UNDEFINED: throw std::runtime_error("");
        case onnx::TensorProto::FLOAT:
Paul's avatar
Paul committed
891
            return literal{{shape::float_type, dims}, t.float_data().begin(), t.float_data().end()};
Paul's avatar
Paul committed
892
893
        case onnx::TensorProto::UINT8: throw std::runtime_error("");
        case onnx::TensorProto::INT8:
Paul's avatar
Paul committed
894
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
895
        case onnx::TensorProto::UINT16:
Paul's avatar
Paul committed
896
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
897
        case onnx::TensorProto::INT16:
Paul's avatar
Paul committed
898
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
899
        case onnx::TensorProto::INT32:
Paul's avatar
Paul committed
900
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
901
        case onnx::TensorProto::INT64:
Paul's avatar
Paul committed
902
            return literal{{shape::int64_type, dims}, t.int64_data().begin(), t.int64_data().end()};
Paul's avatar
Paul committed
903
904
        case onnx::TensorProto::STRING: throw std::runtime_error("");
        case onnx::TensorProto::BOOL:
Paul's avatar
Paul committed
905
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
906
907
        case onnx::TensorProto::FLOAT16:
            return literal{{shape::half_type, dims}, t.float_data().begin(), t.float_data().end()};
Paul's avatar
Paul committed
908
909
910
911
912
913
914
915
        case onnx::TensorProto::DOUBLE:
            return literal{
                {shape::double_type, dims}, t.double_data().begin(), t.double_data().end()};
        case onnx::TensorProto::UINT32: throw std::runtime_error("");
        case onnx::TensorProto::UINT64: throw std::runtime_error("");
        case onnx::TensorProto::COMPLEX64: throw std::runtime_error("");
        case onnx::TensorProto::COMPLEX128: throw std::runtime_error("");
        }
Paul's avatar
Paul committed
916
        MIGRAPHX_THROW("Invalid tensor type");
Paul's avatar
Paul committed
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
    }

    static shape parse_type(const onnx::TypeProto& t)
    {
        shape::type_t shape_type{};
        switch(t.tensor_type().elem_type())
        {
        case onnx::TensorProto::UNDEFINED:
            break; // throw std::runtime_error("Unsupported type UNDEFINED");
        case onnx::TensorProto::FLOAT: shape_type = shape::float_type; break;
        case onnx::TensorProto::UINT8:
            break; // throw std::runtime_error("Unsupported type UINT8");
        case onnx::TensorProto::INT8: shape_type = shape::int8_type; break;
        case onnx::TensorProto::UINT16: shape_type = shape::uint16_type; break;
        case onnx::TensorProto::INT16: shape_type = shape::int16_type; break;
        case onnx::TensorProto::INT32: shape_type = shape::int32_type; break;
        case onnx::TensorProto::INT64: shape_type = shape::int64_type; break;
        case onnx::TensorProto::STRING:
            break; // throw std::runtime_error("Unsupported type STRING");
        case onnx::TensorProto::BOOL:
            break; // throw std::runtime_error("Unsupported type BOOL");
Paul's avatar
Paul committed
938
        case onnx::TensorProto::FLOAT16: shape_type = shape::half_type; break;
Paul's avatar
Paul committed
939
940
941
942
943
944
945
946
947
        case onnx::TensorProto::DOUBLE: shape_type = shape::double_type; break;
        case onnx::TensorProto::UINT32: shape_type = shape::uint32_type; break;
        case onnx::TensorProto::UINT64: shape_type = shape::uint64_type; break;
        case onnx::TensorProto::COMPLEX64:
            break; // throw std::runtime_error("Unsupported type COMPLEX64");
        case onnx::TensorProto::COMPLEX128:
            break; // throw std::runtime_error("Unsupported type COMPLEX128");
        }
        std::vector<std::size_t> dims;
Paul's avatar
Paul committed
948
        auto&& tensor_dims = t.tensor_type().shape().dim();
949
950
951
952
953
954
955
956
957
958
959
        std::transform(tensor_dims.begin(),
                       tensor_dims.end(),
                       std::back_inserter(dims),
                       [](auto&& d) -> std::size_t {
                           if(not d.has_dim_value())
                           {
                               long default_batch_size = 1; // FIXME
                               return default_batch_size;
                           }
                           return d.dim_value();
                       });
Paul's avatar
Paul committed
960
961
        return {shape_type, dims};
    }
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983

    shape::type_t get_type(int dtype)
    {
        switch(dtype)
        {
        case 1: return shape::float_type;
        case 2: return shape::uint8_type;
        case 3: return shape::int8_type;
        case 4: return shape::uint16_type;
        case 5: return shape::int16_type;
        case 6: return shape::int32_type;
        case 7: return shape::int64_type;
        case 10: return shape::half_type;
        case 11: return shape::double_type;
        case 12: return shape::uint32_type;
        case 13: return shape::uint64_type;
        default:
        {
            MIGRAPHX_THROW("Prototensor data type " + std::to_string(dtype) + " not supported");
        }
        }
    }
Paul's avatar
Paul committed
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
};

program parse_onnx(const std::string& name)
{
    std::fstream input(name.c_str(), std::ios::in | std::ios::binary);
    onnx_parser parser;
#ifndef NDEBUG
    // Log the program when it can't be parsed
    try
    {
        parser.parse_from(input);
    }
    catch(...)
    {
        std::cerr << parser.prog << std::endl;
        throw;
    }
#else
    parser.parse_from(input);
#endif
    return std::move(parser.prog);
}

Paul's avatar
Paul committed
1007
} // namespace MIGRAPHX_INLINE_NS
Paul's avatar
Paul committed
1008
} // namespace migraphx